Recently, many videos have appeared on social platforms showing AI editors such as Codex using agents to control industrial software like ANSYS, ABAQUS, and SolidWorks, attracting considerable attention within the community.
Some are excited, saying it can liberate productivity, while others are panicked, because the AI wave has finally broken through the commercial software barrier, and simulation engineers may face unemployment just like frontend programmers.
In this article, we will share our views on the current situation. The following content represents only personal opinions, without referencing AI or translating foreign articles, and is for reference only. Everyone is welcome to discuss passionately in the comments section (please keep language civil).
First, let’s talk about the underlying principles of agent-controlled software operation.
The underlying principles include two aspects:
1. Communication protocol; 2. Software’s own operation API
The prerequisite for an agent to operate commercial software is the need for a secure and reliable communication protocol, meaning the signals sent by the agent can be correctly received by the commercial software.
Taking ANSYS as an example, the current prerequisite for agent operation is to open the MCP server. The core principle of MCP is to convert natural language into the software’s underlying API, and then operate the software through the agent as the hand. But note that the MCP currently on the market is not officially developed by ANSYS and can be downloaded from GitHub.
Therefore, everyone can understand that the so-called software control is fundamentally still about mastering the software’s API, except that compared to the human brain, AI may know more and more comprehensive APIs through learning with the help of software documentation.
Second, let’s discuss the possible AI development directions of commercial software.
Currently, AI’s impact on traditional industries with lower entry barriers such as language, art, and animation is far greater than on industry (no disrespect intended to these industries).
Commercial software, especially strongly industrial software like ANSYS and ABAQUS, is a high ground for AI to conquer. Once this high ground is fully conquered, I personally believe we can say we have entered the AI industrial era.
Facing the AI wave, commercial software companies will certainly develop towards intelligence. This is the future trend. Currently, Shuige believes the possible development directions for commercial software intelligence are:
1. Using general-purpose agent software, with officially maintained interfaces.
General-purpose agents refer to AI programming software that can control the entire local system permissions, such as Codex, Cursor, Trae, etc. Commercial software officials maintain the interface protocols between general-purpose agents and the software, such as the MCP mode mentioned earlier.
Currently, MCP is not officially maintained. Referring to the pyansys project, it may eventually be taken over and maintained by the official team.
2. Commercial software’s own agents
These agents may appear in two forms. One is similar to loading various agent plugins in Visual Studio, achieving natural language interaction with the software’s underlying API through plugins (some ANSYS modules already have AI-driven features). The other type is dedicated agent modules, similar to analysis modules, that can be opened and called separately.
The above two modes are merely Shuige’s guesses. In terms of usage fluency, agents launched by the commercial software itself would certainly be better. However, due to the massive system of commercial software itself, and some modules not yet having underlying APIs, there is still a way to go to achieve smooth natural language interaction with the software.
For more finite element learning resources, please visit Shuige’s blog: FSCAE.COM
Third, AI is reshaping the simulation workflow.
Beginners, after watching videos on various social platforms showing agents operating software, may feel overwhelmed, realizing that software can be used this way. Veterans, after watching, simply silently scroll to the next video.
Admittedly, AI operating software looks very cool, but there are still many problems to be solved. The most core issue is how to make AI truly understand the project rather than just the model during the simulation process.
AI can currently only do work at the numerical level. The so-called numerical level means making cold judgments based solely on numerical values. If it involves more complex project optimization, working condition settings, and other things beyond numerical values, AI cannot yet replace humans.
Even so, AI may be reshaping your simulation workflow. Through continuous learning and optimization, AI will eventually be able to take over the most time-consuming operations in your simulation analysis, such as pre-processing and data extraction.
If you currently only know how to operate the software itself, pursue calculation convergence, mesh generation, and other basic operations, without your own product logic or service advantages, you will most likely be replaced in the future.
Fourth, the cost issue.
Doing anything without talking about cost is just being irresponsible.
The costs most tied to the AI era are time and tokens.
Any sensational effect is simply showing you what they want you to see. Agent-controlled software mainly involves two major costs: time and tokens.
Currently, the videos circulating on the market showing agents operating ANSYS and ABAQUS are mostly carefully edited and accelerated. If you actually go through the operation once, you will find that from natural language instructions to the model truly starting to solve, there are multiple rounds of intent recognition, API matching, and parameter validation in between. If there is any ambiguity, you have to start over.
A mesh that an experienced engineer can finish in three minutes, the agent may struggle for half an hour and still not get it right — those extra 27 minutes are real time costs. In the face of project deadlines, this cost cannot simply be ignored.
Then there is the token cost. Every time an agent initiates an operation, it is essentially calling a large model to generate and execute code. Drawing a mesh, setting a boundary condition, extracting a result — behind the scenes, there may be dozens or even hundreds of API calls. Once it involves large assemblies, complex contacts, or nonlinear materials, the context can easily reach hundreds of millions of tokens. Small companies cannot afford to do this frequently.
For more finite element learning resources, please visit Shuige’s blog: FSCAE.COM
This doesn’t even include the learning cost. To make the agent truly understand your project’s business logic, working condition combinations, and optimization objectives, you need to spend a lot of time feeding context and repeatedly debugging prompts. This itself is a dual expenditure of time and tokens. Many videos that show one-sentence results simply hide all the complex preliminary debugging.
That being said, the technology iteration speed is there, and the cost curve will come down sooner or later. But until then, having a clear understanding of costs will prevent you from being misled by short videos.
Fifth, practical usage limitations.
To make agent-assisted implementation a reality, there are many practical limitations, the most important of which is the security issue.
Using agents for fun is fine for now, but when it comes to actual project simulation, extreme caution is needed.
I don’t know if everyone has the habit of reading privacy policies. For most AI currently available, every word you input is by default used as training material. Only through user-AI interaction can AI continuously learn and evolve in real time.
But if your project involves confidentiality requirements and gets leaked through AI learning, you need to be careful about potential lawsuits!
Sixth, the simulation engineer’s moat in the AI era.
Returning to our initial question, in the current AI era, how should simulation engineers respond?
In Shuige’s view, the AI era has opened up knowledge barriers and amplified individual capabilities. Single simulation operations are becoming increasingly inexpensive. For simulation engineers themselves, to avoid being replaced by AI, we mainly need to improve ourselves in the following three aspects:
1. Return to the problem itself
The essence of simulation is products and services, not software operation.
The software solves the given problem. How to define the problem and how to solve the problem based on the answer is what matters.
Therefore, for engineers, what we need to improve is our understanding of our own products and services, without being too anxious about software operation itself.
For beginners, software operation has become easier to learn. What should mainly be improved is basic simulation knowledge and understanding of underlying principles.
2. Understanding of results
AI has improved the efficiency of single software operations to a certain extent, but are the resulting outcomes reliable? This still requires the engineer’s own ability to judge.
Calculated does not mean correct.
AI only helps engineers accelerate the calculation process, but what truly plays a decisive role is the understanding of the final results.
This understanding includes the engineer’s years of professional experience, knowledge of various working conditions of their own services and products, and more — these are things that AI does not have.
3. Mindset shift: from executor to problem solver
I believe many simulation engineers have a workflow like this: the design department sends a model, you build a finite element model based on it, calculate the model, attach result data and contour plots, obtain key information, feed back to the design department, the design department modifies, and then recalculate…
In such a workflow, the so-called simulation engineer actually only plays the role of a calculator, where both input and output are determined by others.
Combined with the current impact of AI on the industry, the time left for simulation will only decrease in the future. In this situation, appropriately stepping out of the simulation-only circle and trying to transform from executing calculations to solving problems may lead to greater opportunities.
Finally, Shuige believes that you should not reject AI, nor should you race against AI. AI is the future trend. Learn to leverage AI to improve yourself, and spend your time on things only you can do — such as defining the problem truly worth simulating, and making the engineering decisions that concern safety and cost.
Don’t be afraid, and don’t worship AI operating software. Thoroughly understand the product logic and your own services, and AI cannot replace you.
The above represents only my personal views. Everyone is welcome to discuss in the comments section.
At the same time, this article is very welcome to be reposted and shared. Unauthorized direct copying of text to third-party websites is resented.
No comments yet